Re: [PATCH v11 3/9] KVM: VMX: Set host/guest CET states for vmexit/vmentry

[Sean Christopherson]

On Mon, Apr 27, 2020 at 10:04:26AM -0700, Sean Christopherson wrote:
> On Sat, Apr 25, 2020 at 05:20:21PM +0800, Yang Weijiang wrote:
> > On Fri, Apr 24, 2020 at 07:49:41AM -0700, Sean Christopherson wrote:
> > > On Fri, Apr 24, 2020 at 10:35:10PM +0800, Yang Weijiang wrote:
> > > > On Thu, Apr 23, 2020 at 10:17:41AM -0700, Sean Christopherson wrote:
> > > > > On Thu, Mar 26, 2020 at 04:18:40PM +0800, Yang Weijiang wrote:
> > > > > > @@ -7140,8 +7175,23 @@ static void vmx_cpuid_update(struct kvm_vcpu *vcpu)
> > > > > >  	}
> > > > > >  
> > > > > >  	if (guest_cpuid_has(vcpu, X86_FEATURE_SHSTK) ||
> > > > > > -	    guest_cpuid_has(vcpu, X86_FEATURE_IBT))
> > > > > > +	    guest_cpuid_has(vcpu, X86_FEATURE_IBT)) {
> > > > > >  		vmx_update_intercept_for_cet_msr(vcpu);
> > > > > > +
> > > > > > +		if (cpu_has_cet_guest_load_ctrl() && is_cet_supported(vcpu))
> > > > > > +			vm_entry_controls_setbit(to_vmx(vcpu),
> > > > > > +						 VM_ENTRY_LOAD_GUEST_CET_STATE);
> > > > > > +		else
> > > > > > +			vm_entry_controls_clearbit(to_vmx(vcpu),
> > > > > > +						   VM_ENTRY_LOAD_GUEST_CET_STATE);
> > > > > > +
> > > > > > +		if (cpu_has_cet_host_load_ctrl() && is_cet_supported(vcpu))
> > > > > > +			vm_exit_controls_setbit(to_vmx(vcpu),
> > > > > > +						VM_EXIT_LOAD_HOST_CET_STATE);
> > > > > > +		else
> > > > > > +			vm_exit_controls_clearbit(to_vmx(vcpu),
> > > > > > +						  VM_EXIT_LOAD_HOST_CET_STATE);
> > > > > 
> > > > > As above, I think this can be done in vmx_set_cr4().
> > > > >
> > > > Hmm, it's in vmx_set_cr4() in early versions, OK, will move them back.
> > > 
> > > Did I advise you to move them out of vmx_set_cr4()?  It's entirely possible
> > > I forgot some detail since the last time I reviewed this series.
> > Things are always changing, I'm willing to change any part of the patch
> > before it's landed :-).
> 
> I'm worried that there was a reason for requesting the logic to be moved
> out vmx_set_cr4() that I've since forgotten.  I'll see if I can dredge up
> the old mail.

Aha.  v1-v7 had this in cr4.  In v7, you stated that you would move the
toggling to VM-Enter, and in v8 you did just that[2].  In v9 I questioned
why the bits were being toggled in vmx_vcpu_run() and advised moving the
code to vmx_cpuid_update()[3], obviously forgetting that earlier versions
did the toggling in vmx_set_cr4().

AFAICT, I was only reacting to the immediate patch when I advised moving
the code to vmx_cpuid_update(), i.e. the recommendation to move the code
to vmx_set_cr4() doesn't contradict any previous feedback and thus doesn't
reintroduce a known bug.

[1] https://patchwork.kernel.org/patch/11163639/#22931561
[2] https://patchwork.kernel.org/patch/11222763/
[3] https://patchwork.kernel.org/patch/11310823/